AMD: R8xx Speculation

Something boring like:

R600: First DX10 part (from ATI)
Evergreen: First DX11 part.

 

Evergreen appears to be a very substantial evolution of the base architecture it inherited from its predecessors.

The most significant changes to the implementation are parts of DX11 functionality that have not really been analysed, and may not be covered in depth.

Some items might come up whenever the ISA document pops up for the new architecture. Things like what has changed with shared memory, the global synchronization registers, and maybe some GPU-compute related usage of things like the append-consume buffers.
Something like global synchronization registers may be handy, though I've not seen any mention of any other architecture adopting them, which may lead them to be ignored.

The graphics path appears to be a 1 tri/clk setup stage, two rasterizers (contrary to early article statements saying otherwise), no hardware interpolators, two banks of SIMDs, 32 ROPs of some configuration.

I'm not sure if the Evergreen generation will be the last time we have that big read-only crossbar, which has scaled very little since RV770.
Fermi has gone read/write, at some unknown cost. Larrabee has its ring-bus of unknown implementation.
Another question is the future evolution of the caches, and in particular the data paths that feed the SIMDs from the L1s.
In comparison to both Larrabee and Fermi, the number of units that must suck data through that straw is higher.
The read/write loop for the GPU may still be very long, with Fermi somewhere in the middle, and Larrabee with a CPU's short trip to cache.

 

Read only crossbar? Each SIMD has to be able to write to each memory bus ... the crossbar is read/write even if the cache isn't.

 

Shader export has been drawn as something distinct from the crossbar for RV770.

The overview of Cypress doesn't show a change in this regard. The gray lines for the crossbar come from the read-only caches, and the black lines for export go to the ROPs.

 

Shader export essentially is a crossbar though.

 

It seems like it's not the same crossbar, so the loop of read and write traffic still looks to be very long.
Maybe something has changed somewhat with the modification that allows the texture units to read compressed render targets, otherwise the read and write paths don't meet on-chip.

 

Oh shit goes through memory, no doubt ... but there is a write crossbar and it's really not a huge deal to connect it to the cache. All it takes is area.

 

That would be a possible step a later design could take.
The contention for such a cache would be brutal without a read/write cache at the SIMD level to absorb traffic.

 

NDA on 5770 expires 12th or 13th next week – so expect reviews on Tuesday I recon.
Retailers already have stock, waiting for the NDA to pass....

 

Presumably the leaked specs of Juniper cards:
5770: 825 Mhz, 1120 SPs, 56 TMUs, 128 bit, 1 GB GDDR5 1150 Mhz, 149 €
5750: 725 Mhz the rest is the same for 119 €

http://www.fudzilla.com/content/view/15838/1/


How fast do you think 5770/5750 will be compared to 4890/4870?

 

It's just strange we know so little about the 5770 performance and configuration then, especially with that note about retail availability..

Edit: a minute too late
Looks like the '70 will be somewhere between 4870 and 4890 then, probably close to the later when not going too high in msaa and resolution.
Compared to 4770 stock it's 1.9x the shader/texture performance and 1.5x bandwidth.
I remember a review somewhere with a heavy overclocked 4770 going around 4870 performance (when not limited by the 512mb).

 

From here at the end of the article:

Forgive my geography, but is Cozumel even in the northern hemisphere?

 

It honestly saddens me that Fudo let that be published. Bookmark it, you will know why in 3 weeks or less.

-Charlie

 

He probably let them be published because at least german pricewatch lists it as such:
http://geizhals.at/deutschland/a470154.html
Chiptakt: 700MHz, Speichertakt: 1150MHz • Chip: Juniper LE • Speicherinterface: 128-bit • Stream-Prozessoren: 1120 • Textureinheiten: 56 • Fertigung: 40nm • Maximaler Verbrauch: keine Angabe • DirectX: 11 • Shader Modell: 5 • Bauweise: Dual-Slot • Besonderheiten: integrierter 7.1 HD Audiocontroller, unterstützt CrossFireX

 

At this point is very likely that the real Cypress SP count is not 1600...

Anyway yes, the card is in retail and available (someone already bought it in Italy too, this means availabilty is very high, as Italy normally needs more time to get latest tech products in comparison to germany *sigh* ), so far for the "no DX11 in volumes" statement. Also this is a repetition of the "4850 mistake", we have cards and no reviews out....

 

Indeed it is, someone here already calculated it could have up to 24-26 SIMDs in that space (including the TUs of course)

 

So AMD is going 1GB for all products, medium to high end and quite possibly with the same GDDR5 modules?

Do we know anything about the ROPS count?

My guess is that, in crossfire friendly games, two 5750s, could end up faster than the 5850 or even the 5870! I mean, 2X56 Tmus, 2X1120 shaders, 2X128 bit bus, at 240 euros, its getting really interesting!

 

Is it possible that Juniper consists in 14-15 SIMDs ?

Great die size to functional units scaling !

 

Looks pretty consistent:

sapphire 5750: http://www.heise.de/preisvergleich/a470363.html
powercolor 5750: http://www.heise.de/preisvergleich/a470154.html
powercolor 5770: http://www.heise.de/preisvergleich/a470174.html
msi 5770: http://www.heise.de/preisvergleich/a470360.html

 

It has 14 SIMDs